This is the pre-proceedings for the RLC 2026. You may expect minor changes.
Reinforcement Learning Journal, vol. 7, 2026, pp. TBD.
Will be presented at the Reinforcement Learning Conference (RLC), Montréal, Quebec, Canada, August 15–17, 2026.
Keywords: Function approximation, sparse representations, representation learning
Many choices in deep reinforcement learning (DRL) were inherited from deep supervised learning and remain largely unchanged. It is common to use fully connected feedforward neural networks---potentially first with convolutional layers---with rectified linear unit (ReLU) activations, standard initializations, and the Adam optimizer. It is hard to deviate from these defaults because it can be hard to change just one component and maintain performance, and similarly it can be hard to isolate and understand one component. In this work, we reconsider and study one of these choices: replacing ReLU with another activation, called the Fuzzy Tiling Activation (FTA). FTA is a one-to-many activation that was introduced previously to learn sparse activation patterns, and has been shown to be useful for improving stability in online reinforcement learning, continual reinforcement learning, and zero-shot representation transfer. Despite these empirical successes, FTA remains largely underexplored. In this paper, we provide an extensive empirical study of FTA looking at: (1) performance across popular discrete and continuous action benchmarks and four base learning methods, (2) performance on a new environment, (3) stability and reliability across and within runs, (4) robustness to loss of plasticity in continual RL, and (5) hyperparameter sensitivity. Across a large set of experiments we find FTA easy to use and performant. We also contribute improvements to FTA itself including a simple scheme to remove one of its hyperparameters.
Tyler Lazar, Matthew Vandergrift, Martha White, and Adam White. "Revisiting FTA: A Sparse One-to-Many Activation for Reinforcement Learning." Reinforcement Learning Journal, vol. 7, 2026, pp. TBD.
BibTeX:@article{lazar2026revisiting,
title={Revisiting FTA: A Sparse One-to-Many Activation for Reinforcement Learning},
author={Tyler Lazar and Matthew Vandergrift and Martha White and Adam White},
journal={Reinforcement Learning Journal},
volume={7},
pages={},
year={2026}
}